mm/slab_common.c

   1 /*
   2  * Slab allocator functions that are independent of the allocator strategy
   3  *
   4  * (C) 2012 Christoph Lameter <cl@linux.com>
   5  */
   6 #include <linux/slab.h>
   7
   8 #include <linux/mm.h>
   9 #include <linux/poison.h>
  10 #include <linux/interrupt.h>
  11 #include <linux/memory.h>
  12 #include <linux/compiler.h>
  13 #include <linux/module.h>
  14 #include <linux/cpu.h>
  15 #include <linux/uaccess.h>
  16 #include <linux/seq_file.h>
  17 #include <linux/proc_fs.h>
  18 #include <asm/cacheflush.h>
  19 #include <asm/tlbflush.h>
  20 #include <asm/page.h>
  21 #include <linux/memcontrol.h>
  22
  23 #include "slab.h"
  24
  25 enum slab_state slab_state;
  26 LIST_HEAD(slab_caches);
  27 DEFINE_MUTEX(slab_mutex);
  28 struct kmem_cache *kmem_cache;
  29
  30 #ifdef CONFIG_DEBUG_VM
  31 static int kmem_cache_sanity_check(struct mem_cgroup *memcg, const char *name,
  32                                    size_t size)
  33 {
  34         struct kmem_cache *s = NULL;
  35
  36         if (!name || in_interrupt() || size < sizeof(void *) ||
  37                 size > KMALLOC_MAX_SIZE) {
  38                 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
  39                 return -EINVAL;
  40         }
  41
  42         list_for_each_entry(s, &slab_caches, list) {
  43                 char tmp;
  44                 int res;
  45
  46                 /*
  47                  * This happens when the module gets unloaded and doesn't
  48                  * destroy its slab cache and no-one else reuses the vmalloc
  49                  * area of the module.  Print a warning.
  50                  */
  51                 res = probe_kernel_address(s->name, tmp);
  52                 if (res) {
  53                         pr_err("Slab cache with size %d has lost its name\n",
  54                                s->object_size);
  55                         continue;
  56                 }
  57
  58 #if !defined(CONFIG_SLUB)
  59                 /*
  60                  * For simplicity, we won't check this in the list of memcg
  61                  * caches. We have control over memcg naming, and if there
  62                  * aren't duplicates in the global list, there won't be any
  63                  * duplicates in the memcg lists as well.
  64                  */
  65                 if (!memcg && !strcmp(s->name, name)) {
  66                         pr_err("%s (%s): Cache name already exists.\n",
  67                                __func__, name);
  68                         dump_stack();
  69                         s = NULL;
  70                         return -EINVAL;
  71                 }
  72 #endif
  73         }
  74
  75         WARN_ON(strchr(name, ' '));     /* It confuses parsers */
  76         return 0;
  77 }
  78 #else
  79 static inline int kmem_cache_sanity_check(struct mem_cgroup *memcg,
  80                                           const char *name, size_t size)
  81 {
  82         return 0;
  83 }
  84 #endif
  85
  86 #ifdef CONFIG_MEMCG_KMEM
  87 int memcg_update_all_caches(int num_memcgs)
  88 {
  89         struct kmem_cache *s;
  90         int ret = 0;
  91         mutex_lock(&slab_mutex);
  92
  93         list_for_each_entry(s, &slab_caches, list) {
  94                 if (!is_root_cache(s))
  95                         continue;
  96
  97                 ret = memcg_update_cache_size(s, num_memcgs);
  98                 /*
  99                  * See comment in memcontrol.c, memcg_update_cache_size:
 100                  * Instead of freeing the memory, we'll just leave the caches
 101                  * up to this point in an updated state.
 102                  */
 103                 if (ret)
 104                         goto out;
 105         }
 106
 107         memcg_update_array_size(num_memcgs);
 108 out:
 109         mutex_unlock(&slab_mutex);
 110         return ret;
 111 }
 112 #endif
 113
 114 /*
 115  * Figure out what the alignment of the objects will be given a set of
 116  * flags, a user specified alignment and the size of the objects.
 117  */
 118 unsigned long calculate_alignment(unsigned long flags,
 119                 unsigned long align, unsigned long size)
 120 {
 121         /*
 122          * If the user wants hardware cache aligned objects then follow that
 123          * suggestion if the object is sufficiently large.
 124          *
 125          * The hardware cache alignment cannot override the specified
 126          * alignment though. If that is greater then use it.
 127          */
 128         if (flags & SLAB_HWCACHE_ALIGN) {
 129                 unsigned long ralign = cache_line_size();
 130                 while (size <= ralign / 2)
 131                         ralign /= 2;
 132                 align = max(align, ralign);
 133         }
 134
 135         if (align < ARCH_SLAB_MINALIGN)
 136                 align = ARCH_SLAB_MINALIGN;
 137
 138         return ALIGN(align, sizeof(void *));
 139 }
 140
 141
 142 /*
 143  * kmem_cache_create - Create a cache.
 144  * @name: A string which is used in /proc/slabinfo to identify this cache.
 145  * @size: The size of objects to be created in this cache.
 146  * @align: The required alignment for the objects.
 147  * @flags: SLAB flags
 148  * @ctor: A constructor for the objects.
 149  *
 150  * Returns a ptr to the cache on success, NULL on failure.
 151  * Cannot be called within a interrupt, but can be interrupted.
 152  * The @ctor is run when new pages are allocated by the cache.
 153  *
 154  * The flags are
 155  *
 156  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
 157  * to catch references to uninitialised memory.
 158  *
 159  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
 160  * for buffer overruns.
 161  *
 162  * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
 163  * cacheline.  This can be beneficial if you're counting cycles as closely
 164  * as davem.
 165  */
 166
 167 struct kmem_cache *
 168 kmem_cache_create_memcg(struct mem_cgroup *memcg, const char *name, size_t size,
 169                         size_t align, unsigned long flags, void (*ctor)(void *),
 170                         struct kmem_cache *parent_cache)
 171 {
 172         struct kmem_cache *s = NULL;
 173         int err = 0;
 174
 175         get_online_cpus();
 176         mutex_lock(&slab_mutex);
 177
 178         if (!kmem_cache_sanity_check(memcg, name, size) == 0)
 179                 goto out_locked;
 180
 181         /*
 182          * Some allocators will constraint the set of valid flags to a subset
 183          * of all flags. We expect them to define CACHE_CREATE_MASK in this
 184          * case, and we'll just provide them with a sanitized version of the
 185          * passed flags.
 186          */
 187         flags &= CACHE_CREATE_MASK;
 188
 189         s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
 190         if (s)
 191                 goto out_locked;
 192
 193         s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
 194         if (s) {
 195                 s->object_size = s->size = size;
 196                 s->align = calculate_alignment(flags, align, size);
 197                 s->ctor = ctor;
 198
 199                 if (memcg_register_cache(memcg, s, parent_cache)) {
 200                         kmem_cache_free(kmem_cache, s);
 201                         err = -ENOMEM;
 202                         goto out_locked;
 203                 }
 204
 205                 s->name = kstrdup(name, GFP_KERNEL);
 206                 if (!s->name) {
 207                         kmem_cache_free(kmem_cache, s);
 208                         err = -ENOMEM;
 209                         goto out_locked;
 210                 }
 211
 212                 err = __kmem_cache_create(s, flags);
 213                 if (!err) {
 214                         s->refcount = 1;
 215                         list_add(&s->list, &slab_caches);
 216                         memcg_cache_list_add(memcg, s);
 217                 } else {
 218                         kfree(s->name);
 219                         kmem_cache_free(kmem_cache, s);
 220                 }
 221         } else
 222                 err = -ENOMEM;
 223
 224 out_locked:
 225         mutex_unlock(&slab_mutex);
 226         put_online_cpus();
 227
 228         if (err) {
 229
 230                 if (flags & SLAB_PANIC)
 231                         panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
 232                                 name, err);
 233                 else {
 234                         printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
 235                                 name, err);
 236                         dump_stack();
 237                 }
 238
 239                 return NULL;
 240         }
 241
 242         return s;
 243 }
 244
 245 struct kmem_cache *
 246 kmem_cache_create(const char *name, size_t size, size_t align,
 247                   unsigned long flags, void (*ctor)(void *))
 248 {
 249         return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
 250 }
 251 EXPORT_SYMBOL(kmem_cache_create);
 252
 253 void kmem_cache_destroy(struct kmem_cache *s)
 254 {
 255         /* Destroy all the children caches if we aren't a memcg cache */
 256         kmem_cache_destroy_memcg_children(s);
 257
 258         get_online_cpus();
 259         mutex_lock(&slab_mutex);
 260         s->refcount--;
 261         if (!s->refcount) {
 262                 list_del(&s->list);
 263
 264                 if (!__kmem_cache_shutdown(s)) {
 265                         mutex_unlock(&slab_mutex);
 266                         if (s->flags & SLAB_DESTROY_BY_RCU)
 267                                 rcu_barrier();
 268
 269                         memcg_release_cache(s);
 270                         kfree(s->name);
 271                         kmem_cache_free(kmem_cache, s);
 272                 } else {
 273                         list_add(&s->list, &slab_caches);
 274                         mutex_unlock(&slab_mutex);
 275                         printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
 276                                 s->name);
 277                         dump_stack();
 278                 }
 279         } else {
 280                 mutex_unlock(&slab_mutex);
 281         }
 282         put_online_cpus();
 283 }
 284 EXPORT_SYMBOL(kmem_cache_destroy);
 285
 286 int slab_is_available(void)
 287 {
 288         return slab_state >= UP;
 289 }
 290
 291 #ifndef CONFIG_SLOB
 292 /* Create a cache during boot when no slab services are available yet */
 293 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
 294                 unsigned long flags)
 295 {
 296         int err;
 297
 298         s->name = name;
 299         s->size = s->object_size = size;
 300         s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
 301         err = __kmem_cache_create(s, flags);
 302
 303         if (err)
 304                 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
 305                                         name, size, err);
 306
 307         s->refcount = -1;       /* Exempt from merging for now */
 308 }
 309
 310 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
 311                                 unsigned long flags)
 312 {
 313         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
 314
 315         if (!s)
 316                 panic("Out of memory when creating slab %s\n", name);
 317
 318         create_boot_cache(s, name, size, flags);
 319         list_add(&s->list, &slab_caches);
 320         s->refcount = 1;
 321         return s;
 322 }
 323
 324 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
 325 EXPORT_SYMBOL(kmalloc_caches);
 326
 327 #ifdef CONFIG_ZONE_DMA
 328 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
 329 EXPORT_SYMBOL(kmalloc_dma_caches);
 330 #endif
 331
 332 /*
 333  * Conversion table for small slabs sizes / 8 to the index in the
 334  * kmalloc array. This is necessary for slabs < 192 since we have non power
 335  * of two cache sizes there. The size of larger slabs can be determined using
 336  * fls.
 337  */
 338 static s8 size_index[24] = {
 339         3,      /* 8 */
 340         4,      /* 16 */
 341         5,      /* 24 */
 342         5,      /* 32 */
 343         6,      /* 40 */
 344         6,      /* 48 */
 345         6,      /* 56 */
 346         6,      /* 64 */
 347         1,      /* 72 */
 348         1,      /* 80 */
 349         1,      /* 88 */
 350         1,      /* 96 */
 351         7,      /* 104 */
 352         7,      /* 112 */
 353         7,      /* 120 */
 354         7,      /* 128 */
 355         2,      /* 136 */
 356         2,      /* 144 */
 357         2,      /* 152 */
 358         2,      /* 160 */
 359         2,      /* 168 */
 360         2,      /* 176 */
 361         2,      /* 184 */
 362         2       /* 192 */
 363 };
 364
 365 static inline int size_index_elem(size_t bytes)
 366 {
 367         return (bytes - 1) / 8;
 368 }
 369
 370 /*
 371  * Find the kmem_cache structure that serves a given size of
 372  * allocation
 373  */
 374 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
 375 {
 376         int index;
 377
 378         if (size > KMALLOC_MAX_SIZE) {
 379                 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
 380                 return NULL;
 381         }
 382
 383         if (size <= 192) {
 384                 if (!size)
 385                         return ZERO_SIZE_PTR;
 386
 387                 index = size_index[size_index_elem(size)];
 388         } else
 389                 index = fls(size - 1);
 390
 391 #ifdef CONFIG_ZONE_DMA
 392         if (unlikely((flags & GFP_DMA)))
 393                 return kmalloc_dma_caches[index];
 394
 395 #endif
 396         return kmalloc_caches[index];
 397 }
 398
 399 /*
 400  * Create the kmalloc array. Some of the regular kmalloc arrays
 401  * may already have been created because they were needed to
 402  * enable allocations for slab creation.
 403  */
 404 void __init create_kmalloc_caches(unsigned long flags)
 405 {
 406         int i;
 407
 408         /*
 409          * Patch up the size_index table if we have strange large alignment
 410          * requirements for the kmalloc array. This is only the case for
 411          * MIPS it seems. The standard arches will not generate any code here.
 412          *
 413          * Largest permitted alignment is 256 bytes due to the way we
 414          * handle the index determination for the smaller caches.
 415          *
 416          * Make sure that nothing crazy happens if someone starts tinkering
 417          * around with ARCH_KMALLOC_MINALIGN
 418          */
 419         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
 420                 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
 421
 422         for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
 423                 int elem = size_index_elem(i);
 424
 425                 if (elem >= ARRAY_SIZE(size_index))
 426                         break;
 427                 size_index[elem] = KMALLOC_SHIFT_LOW;
 428         }
 429
 430         if (KMALLOC_MIN_SIZE >= 64) {
 431                 /*
 432                  * The 96 byte size cache is not used if the alignment
 433                  * is 64 byte.
 434                  */
 435                 for (i = 64 + 8; i <= 96; i += 8)
 436                         size_index[size_index_elem(i)] = 7;
 437
 438         }
 439
 440         if (KMALLOC_MIN_SIZE >= 128) {
 441                 /*
 442                  * The 192 byte sized cache is not used if the alignment
 443                  * is 128 byte. Redirect kmalloc to use the 256 byte cache
 444                  * instead.
 445                  */
 446                 for (i = 128 + 8; i <= 192; i += 8)
 447                         size_index[size_index_elem(i)] = 8;
 448         }
 449         for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
 450                 if (!kmalloc_caches[i]) {
 451                         kmalloc_caches[i] = create_kmalloc_cache(NULL,
 452                                                         1 << i, flags);
 453                 }
 454
 455                 /*
 456                  * Caches that are not of the two-to-the-power-of size.
 457                  * These have to be created immediately after the
 458                  * earlier power of two caches
 459                  */
 460                 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
 461                         kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
 462
 463                 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
 464                         kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
 465         }
 466
 467         /* Kmalloc array is now usable */
 468         slab_state = UP;
 469
 470         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 471                 struct kmem_cache *s = kmalloc_caches[i];
 472                 char *n;
 473
 474                 if (s) {
 475                         n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
 476
 477                         BUG_ON(!n);
 478                         s->name = n;
 479                 }
 480         }
 481
 482 #ifdef CONFIG_ZONE_DMA
 483         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 484                 struct kmem_cache *s = kmalloc_caches[i];
 485
 486                 if (s) {
 487                         int size = kmalloc_size(i);
 488                         char *n = kasprintf(GFP_NOWAIT,
 489                                  "dma-kmalloc-%d", size);
 490
 491                         BUG_ON(!n);
 492                         kmalloc_dma_caches[i] = create_kmalloc_cache(n,
 493                                 size, SLAB_CACHE_DMA | flags);
 494                 }
 495         }
 496 #endif
 497 }
 498 #endif /* !CONFIG_SLOB */
 499
 500
 501 #ifdef CONFIG_SLABINFO
 502 void print_slabinfo_header(struct seq_file *m)
 503 {
 504         /*
 505          * Output format version, so at least we can change it
 506          * without _too_ many complaints.
 507          */
 508 #ifdef CONFIG_DEBUG_SLAB
 509         seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
 510 #else
 511         seq_puts(m, "slabinfo - version: 2.1\n");
 512 #endif
 513         seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
 514                  "<objperslab> <pagesperslab>");
 515         seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
 516         seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
 517 #ifdef CONFIG_DEBUG_SLAB
 518         seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
 519                  "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
 520         seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
 521 #endif
 522         seq_putc(m, '\n');
 523 }
 524
 525 static void *s_start(struct seq_file *m, loff_t *pos)
 526 {
 527         loff_t n = *pos;
 528
 529         mutex_lock(&slab_mutex);
 530         if (!n)
 531                 print_slabinfo_header(m);
 532
 533         return seq_list_start(&slab_caches, *pos);
 534 }
 535
 536 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
 537 {
 538         return seq_list_next(p, &slab_caches, pos);
 539 }
 540
 541 static void s_stop(struct seq_file *m, void *p)
 542 {
 543         mutex_unlock(&slab_mutex);
 544 }
 545
 546 static void
 547 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
 548 {
 549         struct kmem_cache *c;
 550         struct slabinfo sinfo;
 551         int i;
 552
 553         if (!is_root_cache(s))
 554                 return;
 555
 556         for_each_memcg_cache_index(i) {
 557                 c = cache_from_memcg(s, i);
 558                 if (!c)
 559                         continue;
 560
 561                 memset(&sinfo, 0, sizeof(sinfo));
 562                 get_slabinfo(c, &sinfo);
 563
 564                 info->active_slabs += sinfo.active_slabs;
 565                 info->num_slabs += sinfo.num_slabs;
 566                 info->shared_avail += sinfo.shared_avail;
 567                 info->active_objs += sinfo.active_objs;
 568                 info->num_objs += sinfo.num_objs;
 569         }
 570 }
 571
 572 int cache_show(struct kmem_cache *s, struct seq_file *m)
 573 {
 574         struct slabinfo sinfo;
 575
 576         memset(&sinfo, 0, sizeof(sinfo));
 577         get_slabinfo(s, &sinfo);
 578
 579         memcg_accumulate_slabinfo(s, &sinfo);
 580
 581         seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
 582                    cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
 583                    sinfo.objects_per_slab, (1 << sinfo.cache_order));
 584
 585         seq_printf(m, " : tunables %4u %4u %4u",
 586                    sinfo.limit, sinfo.batchcount, sinfo.shared);
 587         seq_printf(m, " : slabdata %6lu %6lu %6lu",
 588                    sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
 589         slabinfo_show_stats(m, s);
 590         seq_putc(m, '\n');
 591         return 0;
 592 }
 593
 594 static int s_show(struct seq_file *m, void *p)
 595 {
 596         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
 597
 598         if (!is_root_cache(s))
 599                 return 0;
 600         return cache_show(s, m);
 601 }
 602
 603 /*
 604  * slabinfo_op - iterator that generates /proc/slabinfo
 605  *
 606  * Output layout:
 607  * cache-name
 608  * num-active-objs
 609  * total-objs
 610  * object size
 611  * num-active-slabs
 612  * total-slabs
 613  * num-pages-per-slab
 614  * + further values on SMP and with statistics enabled
 615  */
 616 static const struct seq_operations slabinfo_op = {
 617         .start = s_start,
 618         .next = s_next,
 619         .stop = s_stop,
 620         .show = s_show,
 621 };
 622
 623 static int slabinfo_open(struct inode *inode, struct file *file)
 624 {
 625         return seq_open(file, &slabinfo_op);
 626 }
 627
 628 static const struct file_operations proc_slabinfo_operations = {
 629         .open           = slabinfo_open,
 630         .read           = seq_read,
 631         .write          = slabinfo_write,
 632         .llseek         = seq_lseek,
 633         .release        = seq_release,
 634 };
 635
 636 static int __init slab_proc_init(void)
 637 {
 638         proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
 639         return 0;
 640 }
 641 module_init(slab_proc_init);
 642 #endif /* CONFIG_SLABINFO */